Article title: Siamese neural networks for the classification of high-dimensional radiomic features

A Siamese neural network contains at least two identical subnetworks (same configuration, parameters and weights) and its purpose is to determine the similarity between the inputs (one per subnetwork). Traditionally, neural networks require a large amount of data to work effectively, which is quite rare to find in highly specific medical image analysis problems.

This study’s purpose was to demonstrate the effectiveness of a Siamese neural network at classifying numerous features (ex : shape and size-based features) extracted from MRI scans of a small sample of female patients with breast cancer (55 subjects) or granulomatous lobular mastitis (GLM) (44 subjects), an inflammatory breast condition.

First, to localize the region of interest and then extract the features, the lesions caused by each pathology were segmented by an experienced radiologist. Afterwards, the most significant features according to a Student T-Test were used to train and evaluate the model. The test features were associated to the same pathology as the training features with whom they shared the highest network-given similarity after pairing them. The features were separated in four groups and the results obtained with the Siamese network and those from two conventional non-neural-network models (Discriminant Analysis (DA) and Support Vector Machine (SVM)) are shown in Table 1.

Table 1 – Results from using the various features and models to distinguish the two pathologies.

According to Table 1, when it comes to the two biggest feature sets, the Siamese network outperforms the two other models on area-under-curve (AUC), accuracy, sensitivity, and specificity metrics. However, when it comes to the two smallest feature sets, it performed worse. It appears this type of neural network suffers in lower-dimensional spaces compared to higher-dimensional spaces.

Reference: Mahajan, A., Dormer, J., Li, Q., Chen, D., Zhang, Z., & Fei, B. (2020). Siamese neural networks for the classification of high-dimensional radiomic features. Proceedings of SPIE–the International Society for Optical Engineering, 11314, 113143Q. https://doi.org/10.1117/12.2549389